SP Arun

Come join us for this exciting event next week. You can't possibly have anything better to do next Saturday 🙃

Maybe so, if all edges are detected nicely. Computer vision based edge detection doesn't look like this.

We are grateful to the (late) Carl Olson, @nancykanwisher.bsky.social and many others for their critical and constructive inputs, to the awesome India Alliance for the blue-sky funding and to CNS & IISc for continuing to be an awesome place to work! And we're done! Thanks for making it this far! n/n

(started in Aug 2022, got reviewed by several high profile journals, got kicked out each time due to 1 of 3 dissenting reviewers who gave single-paragraph reviews, etc., and finally ended up with eLife where the reviewers still gave us a hard time but we are done) 29/n

For now we are content with showing a really cool set of results, on which we have worked hard over several years, and shepherded through an unusually long review process 28/n

We think this computation could predict all kinds of visual behaviours – such as visual search asymmetry, aesthetic appeal, perceived complexity and image memorability. All this needs to be investigated thoroughly before we know what this all means. 27/n

To sum up, we think we’ve found experimental evidence that there’s a novel computation called visual homogeneity, being performed in a localized region in the brain, which is being used to solve property-based visual tasks. 26/n

If visual homogeneity is a really universal computation, then it should also work for a symmetry task. Here too, we obtained exactly similar results - the visual homogeneity computation predicted symmetry responses, and the same region showed proportional activations!! 25/n

This region is just anterior to the lateral occipital complex, where neural dissimilarity between images matched perceived dissimilarity 24/n

In the brain, we found a localized region whose brain activity is directly proportional to visual homogeneity. 23/n

So armed with these predictions, we collected and analyzed our data.....and lo and behold! We found that indeed, exactly as we predicted, we can find a center in perceptual space relative to which distance computations do predict oddball present/absent search 22/n

By contrast, regions that encode task difficulty will not show this pattern, because task difficulty is largest in the middle range of VH and not the extremes 21/n

In brain activity, if there's a single brain region (region VH) that encodes this quantity, the response in this region should be directly proportional to visual homogeneity. 20/n

If such a computation was actually being used by the brain, what do we expect? Well, first of all, it should act like a decision variable, which means that any stimulus close to the decision boundary will be hard to decide, and will have long response times. 19/n

Because this is a very specific computation that involves calculating the distance of each visual stimulus to some hypothetical center in neural space, we called this quantity/computation as "visual homogeneity". 18/n

The same idea would work for symmetry tasks: because the symmetric object has the same visual features repeated, it will "stand apart" compared to asymmetric or visually heterogeneous arrays. Thus we could "solve" a symmetry task by computing the distance to some center 17/n

What this means is that visually homogeneous arrays would automatically create a neural response that "stands apart" compared to visually heterogeneous arrays. Thus we could "solve" the oddball search task by simply computing the distance of each image to some center 16/n

So, when you see an array containing identical items the neural response is equal to the single item response (well, almost, to an approximation). But when you see an array containing an oddball, its representation is somewhere between the neural response to the two items. 15/n

Well to test it, first of all, we need to explain how we get the response to a multi-item array from the response to a single item. Luckily, we know the answer: the neural response is the average of the single item responses, as shown by us and others 14/n

YouTube video by Nick Stöpler Feynman: It doesn't matter how beautiful your theory is, it doesn't matter how smart you are..

Well in science, just having a cool idea isn't enough, you got to show that it works. And as Feynman says, if it doesn't work, it doesn't matter how famous or rich you are, its just wrong. Its a heartbreaking but beautiful aspect of doing science. 13/n
www.youtube.com/watch?v=p2xh...

We thought maybe we use the same underlying computation to solve all these three tasks, despite the fact that our verbal description for them is entirely different! Cool no? 12/n

In a same-different task, the "same" display is visually homogeneous. In an oddball search task, the "target-absent" array is visually homogeneous. In a symmetry task, the symmetric object is visually homogeneous. 11/n

We realized that the common property that across all these property-based tasks (same-different, oddball search and symmetry) is that you are trying to discriminate a visually homogeneous image from a visually heterogeneous image 10/n

Our key insight was based on our earlier work where we showed that symmetric objects become special due to simple computations in neurons, and that perceiving symmetry does not require "symmetry detectors" as many have proposed 9/n

What's the feature space used by the brain to solve these tasks? What's the decision variable? You cannot solve these tasks by looking for any particular feature in feature space, because an object can have any set of features and be symmetric or asymmetric. 8/n

But there's an entire other category of visual tasks that do not fall into this framework. These tasks involve searching for a particular property, like determining if two images are same or different, deciding if there's an odd one out, and even deciding symmetry! 7/n

You see, most visual tasks are feature-based. If you are searching for the handsome Georgin in a crowd while ignoring other ugly distractors, you'd have to train a classifier, and project any image to get a decision variable. This is the standard model for decision making 6/n

If there's such systematic variation, there's some intrinsic image property that makes it a "good" or "bad" distractor. What could this be? For a while we puzzled over this question, but we quickly noticed an even more fundamental question. How do we even solve this task?! 5/n

Don't believe it? Below are some example arrays in which you should confirm if there is or isn't an oddball. You might find that its easy to confirm that there's no target in the rope array (mean RT = 1.06 s) than in the leaf array (mean RT = 1.45 s). Why?!! 4/n

Now this posed a puzzle - when there is an oddball target among distractors, its well known that its easy to find if it is dissimilar to the distractors. But in target-absent searches, there is no target, so why would you take more or less time to "not-find" a target? 3/n